AUDR at Image CLEF 2011: Medical retrieval task

This paper describes the participation of AUDR group in the ImageCLEF 2011 medical retrieval task. We are particularly interested in the adhoc retrieval task. Two existing retrieval engines are used: LIRE for visual retrieval and Apache Lucene for textual retrieval. Based on the two tools, we consider two information models for text: vector space model and topic model, and the query expasion mechansim based on KL distance is used for improving the performance. Concerning the visual retrieval, a purely visual approach named CEDD is used with LIRE. Fusion strategies are also tried out to combine results from two engines. The experiments carried out on the ImageCLEFmed datasets show baselines provided by LIRE and Lucene are ranked close to the average among visual and textual runs respectively, and the fusion of vector spacing model and topic model will perform better than mono-model.